Consolidates all GPU device memory allocation under a single VMM-backed allocator, removing manual memory fraction settings and the dependency on cudaMemAlloc, which keeps memory addresses stable across CUDA graph iterations.
Unifies GPU device memory allocation under DeviceAddressVmmAllocator with NCCL symmetric memory compatibility and a rank-symmetric reuse partition to prevent bootstrap-barrier deadlocks in multi-device training.
Large-scale multi-GPU training runs can avoid crashes caused by inconsistent memory addresses breaking CUDA graphs and by rank-asymmetric memory pools that trigger bootstrap-barrier deadlocks, so operators running SPMD workloads on multiple devices no longer have to tune manual memory fractions or retry failed AllReduce synchronizations. The VMM-backed allocator also opens a path to tighter NCCL symmetric-buffer usage, but teams should watch for regressions on single-GPU inference and smaller workloads that may not exercise the new partition logic, and monitor whether the thread-local DeviceAssignmentScope guard introduces subtle state leaks in long-running services.